AI Proved Signal's Encryption Secure in 4 Weeks — Work That Used to Take Years

A black combination padlock placed on a laptop computer

Photo: Hippopx (public domain). Illustrative image.

What happened. Google DeepMind announced AlphaProof Code on 10 November 2026, a multi-LLM agentic framework that formally verified key security properties of the actual code underlying the Signal protocol — the encryption that protects billions of conversations in Signal and Google Messages. Formal verification means proving, with mathematical rigour, that code does what it claims; it is the strongest guarantee in software security, and historically one of the slowest to produce. DeepMind says the verification took 4 weeks, against work that has historically taken expert teams years.

How it worked. Agents powered by Gemini Carbon generated over 6 million lines of Lean proof code — Lean being the proof language in which the verification is written and machine-checked. (DeepMind’s own blog description says “over three million lines of machine-checked proof”; the technical paper puts the figure at about 6.5 million lines. We use the paper’s figure, with the discrepancy noted.) Humans were not cut out of the loop: experts reviewed every specification, and the agents flagged only about 1% of the codebase for a second human look — a division of labour in which the AI does the volume and humans keep the judgement.

The control test is the convincing part. To check the system actually catches problems rather than rubber-stamping code, DeepMind reintroduced two real, previously fixed bugs into an isolated copy of Signal’s code. In both cases, the system produced a machine-checked proof that the buggy version was insecure. A verification system that passes everything, including broken code, is worthless; this one failed the broken code, correctly. The formal verification artifacts are being open-sourced, and a technical paper was published alongside the announcement.

Why it matters. Signal is about as high-stakes a target as exists: its protocol secures the messages of billions of people, and its code is among the most scrutinised in the world. If AI can formally verify production cryptography at this scale in weeks rather than years, the economics of proving software correct change — and the same techniques could extend to the infrastructure code that everything else depends on. It is also a different kind of AI safety story from the usual alignment debate: not a model refusing bad requests, but a model producing a checkable proof. The same week, Cohere launched Cohere Cyber as a security business, and OpenAI’s Daybreak project set out to automate red-teaming. For the background, see What Is AI Safety?.

Sources

  • AlphaProof Code
  • Google DeepMind
  • Signal
  • formal verification
  • AI safety
  • Lean

Related articles